Perception visuelle computationnelle et cadre d'apprentissage pour l'imagerie médicale axée sur la qualité

Computational Visual Perception and Learning framework for Quality-Driven Medical Imaging

• Auteur / Autrice : Yousra Nabila Taifour
• Direction : Azeddine Beghdadi, Marie Luong, Ming Zuheng
• Type : Projet de thèse
• Discipline(s) : Informatique
• Date : Inscription en doctorat le 19/10/2023
• Etablissement(s) : Paris 13
• Ecole(s) doctorale(s) : École doctorale Galilée (Villetaneuse, Seine-Saint-Denis)

Résumé

Medical imaging and guided cancer treatments heavily rely on precise and certain diagnoses. The integration of deep learning and intelligent solutions into these diagonosis holds significant promise.However, a noteworthy challenge lies in the current lack of explainability and transparency. Another concern is that the gradual changes in treated organs necessitate vigilant treatment and follow-up, exposing patients to potential risks of secondary cancers due to persistent X-ray exposure.A promissing remedy is to minimize the radiation exposure dose for acquiring CT and PET scans. However, this process introduces noise and artifacts in low-dose computed tomography (LDCT) and low-dose positron emission tomography (LDPET) images. Addressing these issues, our research focuses on three critical directions: firstly, developing straightforward methods using deep learning to enhance image quality in low-dose medical images, particularly by considering the modeling of retino-cortical mechanisms; secondly, creating tools that can automatically identify and detect cancerous areas in images, ensuring transparency in our approaches; and thirdly, establishing clear metrics to assess image quality at different stages. The introduction of transparency to the network is considered essential in the medical field, ensuring a clear understanding of the decision-making process and fostering trust in the system. By combining simple ways of modeling how we perceive things, especially from a retinal perspective, with machine learning, our aim is to provide practical and understandable solutions for real issues in medical imaging and diagnosis. This is particularly relevant with the increasing use of Low Dose protocols in cancer treatment.